Leaked ChatGPT debug stream reveals one prompt triggers 18 hidden queries and 50 engine calls
metehan777 · x · 2026-09-08
Developer metehan777 discovered that ChatGPT's server-sent events stream carries a full debug view of its web retrieval pipeline, and spent days mining it (getting rate-limited by Saturday):
- One question is never one search: for a single "best AI visibility tools" prompt, ChatGPT ran 5 search rounds, wrote 18 hidden queries, made 50 engine calls, pulled 228 results, fetched 223 URLs in chunks, and cited only 16 links.
- Pages are rendered as markdown-like text: fingerprinting suggests a customized Python html2text build. Pages are split into 170-word blocks and scored; the model sees only 1-3 blocks per source, never the full page. The ogdata object is always empty; raw metatags are what's kept.
- The same stream also carries: per-query routing to 10 internal engines (web, news, Wikipedia, Reddit, arXiv, YouTube, PDF...), result-level and per-chunk score objects, shouldfetch decisions with crawl/publication dates, a second in-reasoning ranking pass that re-orders domains, the exact model prompt with per-source word budgets, and separate field sets for shopping and local queries.
A single prompt produces a 150,000-line JSON dump. The author calls it the closest look yet at ChatGPT's retrieval pipeline for anyone working on GEO/AI visibility, with more to come this week.
Related event: Dev uncovers ChatGPT's full search debug data in SSE stream(2 posts)→
More from Models
- Kimi K3.1 Pro tops MathArena open models, closing frontier gap to about 6 months — teortaxesTex · 2026-10-26
- Ruff creator Charlie Marsh: frontier models find infinite bugs in any software — charliermarsh · 2026-10-26
- Dev testing one skill across Opus 5.5 and Sol 6.1 finds instruction following 'all over the map' — lucasmeijer · 2026-10-26
- supermemory unveils learner-1, claiming continual learning for any model — julianweisser · 2026-10-26
- Anthropic says Opus 5.5 cuts cost per task by ~40% vs Opus 5 — taufiqintech · 2026-10-26
- AI search APIs ranked: top 12 within 7 points, yet prices differ nearly 7x — FinanceYF5 · 2026-10-26